This repository provides code for implementing the "weak-to-strong generalization" technique, enabling the training of powerful models using labels generated by weaker, less capable models. It's designed for researchers and practitioners in machine learning, particularly those working with large language models and computer vision tasks, to improve model performance and data efficiency.

How It Works

The core approach involves fine-tuning a strong model using labels derived from a weaker model, potentially with auxiliary losses like confidence weighting. The sweep.py script orchestrates this by first training ground truth models for specified sizes and then iteratively training stronger models using the labels from weaker ones. This method aims to transfer knowledge effectively, reducing the need for extensive human-labeled data for high-performance models.

Quick Start & Requirements

• Install dependencies using pip: pip install .
• Requires Python.
• See notebooks/Plotting.ipynb for plotting results.

Highlighted Details

• Implements weak-to-strong learning for binary classification tasks.
• Supports fine-tuning pretrained language models and training against model-generated labels.
• Includes code for weak-to-strong generalization in vision models (AlexNet -> DINO on ImageNet).
• Offers various loss functions described in the paper, including confidence auxiliary loss.

Maintenance & Community

• Authors include Adrien Ecoffet, Manas Joglekar, Jeffrey Wu, Jan Hendrik Kirchner, and Pavel Izmailov (vision).
• Acknowledges Hugging Face for their transformer models.

Licensing & Compatibility

• Licensed under the MIT License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The codebase is noted as not well-tested and may not use the exact settings from the paper, though it aims for qualitatively similar results.